- Relax return value requirements on errors. Only the error is checked,
as there are multiple conflicting return styles throughout the `os`
API.
Some return the error along with `0`, `-1`, or another value. This can
be smoothed out later.
- Test `os2` files now.
- No longer test streams after closing; this is considered similar to a
use-after-free in `os2`.
`open` specifies the `mode` argument as vararg (presumably to make it
optional). varargs actually have rules about casting, in this case the
rule that any integer arg of size <= 4 has to be casted to `i32` before
passing it.
Not doing that implicit cast makes the permissions wrong or not apply at
all.
This new algorithm uses a Scalar->Vector->Scalar iteration loop which
requires no masking off of any incomplete data chunks.
Also, the width was reduced to 32 bytes instead of 64, as I found this
to be about as fast as the previous 64-byte x86 version.
- "null" is the proper way to represent the nil value in the diagnostic
format
- hex encoding in diagnostic format was wrong
- struct keys weren't sorted the right deterministic way
Adds `expect_leak_or_bad_free :: proc(t: ^T, client_test: proc(t: ^T), verifier: Memory_Verifier_Proc)`.
It sets up its own `Tracking_Allocator`, runs the `client_test`, and then calls the `verifier` procedure.
The verifier can then inspect the contents of the tracking allocator and call `testing.expect*` as sensible for the test in question.
Any allocations are then cleared so that the test runner doesn't itself complain about leaks.
Additionally, `ODIN_TEST_LOG_LEVEL_MEMORY` has been added as a define to set the severity of the test runner's memory tracker. You can use `-define:ODIN_TEST_LOG_LEVEL_MEMORY=error` to make tests fail rather than warn if leaks or bad frees have been found.
